Facts are extracted per turn and written to a persistent store rather than re-read from context. We report recall at 7, 30 and 90 days, the rate of false memories, and what happens when a user contradicts something they said a month earlier.
Most conversational systems simulate memory by re-reading a transcript: the illusion holds only as long as the whole exchange fits in context, and it resets the moment it does not. Pelagea-1 does something narrower and more durable. A secondary extraction pass runs over every turn, pulls out discrete, timestamped facts, and writes them to a per-user store that persists independently of any single conversation. The model does not remember the exchange. It remembers what the exchange was about.
The cohort for this note ran 41 users over 90 days, contributing the 12.4k annotated human turns referenced across our other notes. Each extracted fact was hand-verified against the source turn by a second annotator, giving us a clean baseline to measure recall against rather than trusting the model's own account of what it retained.
Recall held at 96.8% after 7 days, 94.1% after 30, and 87.3% after 90. The decay is not uniform: facts tied to a specific, nameable entity (a sister's name, a commute route) held far better than facts tied to a transient state (how someone said they felt on a given day). False memories, meaning facts the model asserted but that were never actually stated, occurred at a rate of 1.2 per 1,000 turns, almost always by conflating two separate users' near-identical statements during an earlier, less isolated version of the extraction pipeline. That failure mode is closed in the current build.
Contradiction is the harder case. When a user restates a fact differently a month later, the store does not silently overwrite. Both versions are kept, timestamped, and the model is instructed to defer to the more recent one while remaining able to surface the earlier version if asked directly. In practice this reads less like a database update and more like the way a person revises what they thought they knew, which was the point: memory that only ever accumulates, never corrects, stops resembling anyone.